Ricursive AI unveils Ripple-65B, a dense open-weight model achieving top scores on terminal-bench 2.1

Ricursive Intelligence has released Ripple-65B Preview, a 65 billion parameter dense, open-weight model engineered specifically for agentic coding. This model achieves a score of 80.52 on Terminal-Bench 2.1, which Ricursive claims is the highest among US open-weight models, surpassing MoE rivals that possess between 501B and 975B parameters. The release package includes three checkpoints: a BF16 instruction-tuned version, an NVFP4 variant, and a BF16 base model, all released under the Apache 2.0 license. The development cost for Ricursive was reported at $32 million, which the company suggests is one-tenth the expense of the previous open-weight agentic coding leader. The architecture features a 5120 d_model across 80 layers, adopting a design similar to Gemma 3N for stability. Training involved a text-only corpus, primarily English and code, at an 8K sequence length. The model's strong performance on Terminal-Bench 2.1 was achieved using only 1.4 million rollouts, significantly fewer than the over 100 million used by Beam. Furthermore, when tested on the official Terminus 2 harness at 128K, Beam's score dropped to 70.8, while Muse Spark Open reached 78.7, remaining below Ripple's 80.5. The model is available on Hugging Face, with initial distribution partners including Together AI, GMI Cloud, and OpenRouter.